https://www.irjet.net/archives/V6/i5/IRJET-V6I51192.pdf

1. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 05 | May 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 8053
Android Malware Detection using Deep Learning
Devi K.R1
Student, Dept. of CSE, College of Engineering Trivandrum, Kerala, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Mobile devices are prone to malware attacks.
Many systems have been implemented to prevent these
attacks but none are fruitful. The implemented system is a
machine learning based malware detection framework
which is used to protect the Android devices from major
security threats. A large collection of dataset is used for
training from which requested permissions are extracted.
Based on these extracted permissions, a model is developed
using the dataset and is tested using unknown malware and
benign app samples.
Key Words: Machine Learning, Malware detection,
Permissions
1. INTRODUCTION
Malware is a software which can cause potential threats to a
computer, server, client, or computer network. Malware
causes damage after it is implanted or introduced into a
target's computer and it is in the form of an executable
codes, script files and other softwares. The codes are known
as viruses, ransomware, spyware, adware, worms , Trojans
,scareware and many other forms. The commonly used
methods for protecting against malware is to prevent the
software from gainingaccesstothetargetcomputerincludes
antivirus software, firewallsandmanyothertechniques. The
main uses of them include preventing their access to target
computers, checking the presence of suspicious activities,
recover from malware attacks. Another strategy to
differentiate malware apps from genuine Android apps
includes sophisticated dynamic and static analysis tools to
detect and classify malicious apps automatically. There are
encryption techniques which will decrease the chances of
malwares from being detected. To avoid this problem, we
can study Android apps to extract permissions which are
sensitive that are widely used in Android malwares. An
automated malware detection system isusedtofightagainst
malwares and assist Android appmarketplacestodetectand
remove unknown malicious apps.
Static analysis tools are used to extract source codes or byte
codes, often traversing the paths of programs to check for
some unique and hidden resources. Static analysis
approaches are used for different tasks which includes the
behaviour assessment of Android apps, detection of
application clones, automatic test case generations, or for
uncovering non functional issues related to performance.
The important point which is to be noted is that the code is
not executed or run but the tool itself isexecuted.Thesource
code is the input to the tool and the mined features are the
output.eg:-Drebin
Dynamic program analysis is the analysis of Android
applications by executing the programs on a virtual
environment like Android Studio. The target programsmust
be executed with test inputs toproducethebehavior.System
calls are analyzed to monitor the behaviour of Android
applications.eg:-TaintDroid
Malware classification is an open problem commonly
rectified by employing machine learning techniques.
Permissions and API calls are extracted w Man is able to
detect behaviors which are sensitive from Android
applications. Most of the detections are based on the
difference of permissions detected by benign apps and
malware apps. By analysing the permissions requested and
api call usages, benign app and malware app samples can
effectively expose abnormal behaviors and finally
distinguish malware from many genuine applications.
So considering the drawbacks of the above techniques we
propose a new model which is based on the extracted
permissions from the apks and uses deep learning
techniques to formulate the model.
2. SYSTEM DESIGN
Most of the malware detection tools uses the manual of lists
of features based on permissions, api calls, sensitive
resources, intents, etc., which are difficult to come by. To
address this problem, we study the different real Android
applications to mine hidden patterns of malware and are
able to extract highly sensitive permissions that are widely
used in Android malware.
Benign apps are downloaded from apkpure.com which is a
free site of benign apks from google playstore. Malicious
apps are downloadedandareextractedfromvirusShare.com
and Contagio Mini Dump. Features like Api related
Permissions are considered to develop the system.
Permission Distribution
Permissions[1] from malwares and benign apps are
identified. By analysis, Access_wifi_state,SendSms etc are
commonly used by malwares. The requested permissions of
the android applications are declared in a file calledAndroid
manifest of the respective apks. From the manifest files,
permissions are extracted and are converted to a csv file. A
large number of permissions are identified in the previous
step. Out of which a few must be selected for further

2. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 05 | May 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 8054
processing. For that Mann Whitney test[2] is employed. For
each permission, if a particular app uses that permission,the
corresponding permission is set as 1 or else it is set to 0.
These values are indicatedbypvalues.[2]Thereforetwosets
of samples are generated. One to represent one specific
permission usage of malicious apps and the other to
represent specific permission usage of benign apps. In the
previously created input file, a comparison test is applied.
For each permission, the average values are computed for
each of the feature vector. So from two sets of samples, we
compute the average values. And those permissions with
higher average values will be selected as the feature vector
for training.
2.1 Malware Detection
This feature vector is divided into two. One can be used for
training the model and the othercanbeusedfordetermining
the model parameters. The first feature vector is fed to the
classifier. The classifier employed here is the Neural
Networks and K-Means Clustering Algorithm. Two trained
models will be created. The second feature vector is given as
input to the model to determine the model parameters like
accuracy, precision, recall etc. Unknown apks are then given
as input to the model so that the model will predict these
apks as benign or malicious.
3. IMPLEMENTATION
Here, we take a closer look at how the system was
implemented. The whole system was developed using
python language.
Benign apps are downloaded from apkpure.com. Malicious
apps are downloaded and extracted from virusshare.com
and Contagio Minidump. A total of 135 benign apps were
collected. A total of 327 malicious apps were collected. The
features namely permissions are extracted using Python 3.7
in Spyder. A package called Androguard[5] is used toextract
manifest files from apks. The extracted permissions are
correctly displayed on the screen. Feature Selection is done
using Extra Tree Classifier which is included as a built in
package in python. Feature selection is performed
successfully using the dataset.
Feature Vectors are generated by using Mann-Whitney
test[3]. It is implemented using the inbuilt package called
scipy.stats in Python 3.7. The weights and their
corresponding feature names are written to a csv file.
Training phase receives a training dataset which is a csv file.
The model is trained using Neural networks and k-means
clustering algorithm. Output of this phase is a confusion
matrix and graphs showing the dataset which are classified
correctly and incorrectly. The model is testedusingdifferent
samples of both malware and benign apps. The output ofthe
feature map as well as the prediction will be printed on the
screen.
The feature map generated for the training data sample is
given in the figure below:-
Figure 1: Feature Map of the training dataset
The extracted permissions from the testing data sample is
shown in the figure below:-
Figure 2: Extracted Features from the test dataset
The feature map of the test data sample and the prediction
is shown in the figure below:-
Figure 3: Feature Map of test data
4. PERFORMANCE ANALYSIS
Performance Analysis deals with the measurement of
response time, Correctness of output and throughput of the
proposed software.

3. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 05 | May 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 8055
A confusion matrix[7] is a table which is used to describethe
performance of a classification model when a set of test data
values are given as input to the trained system.Theaccuracy
of the system gives us an overview of how accurate the
system is when test samples are passed through it. The
accuracy of the model is 88%. The accuracy is less because
this system will bypass malwares using encryption
techniques[4] and java reflection to encrypt source codes.
This system is vulnerable to pollution attacks[5] which
means malwaresrequestpermissionsnormallyrequested by
benign app samples to avoid detection. The accuracy of the
system is shown in the figure below.
Figure 4: Confusion Matrix and accuracy
When compared to other techniques, this accuracy rate is
very high. This model produces a good result when
compared to Drebin[2], Taintdroid[3] and other Machine
Learning techniques.
5. CONCLUSIONS
The implemented system collects datas in the form of
Android apks from various Internet sources. The apks are
extracted to collect features whicharebasicallypermissions.
A feature vector is created based on the permissions andthe
given apks. This is the input to the ML algorithms to build a
trained model. Unknown applications are used as input. An
overall accuracy of 88 percent is achieved.
The main limitations of the model include:-
 A large dataset must be collected to avoid
overfitting problem.
 The extracted permissions are limited because the
number of malicious applications on the internet
are very less.
 This system considers the differences of malware
and benign apps but it does not consider the
categories of benign apps which can be useful for
malware detection.
 This system is open to Mimicry and Pollution
attacks.
 This system bypass malwares using Java Reection
and bytecode encryption.
REFERENCES
[1] G. Tao, Z. Zheng, Z. Guo and M. R. Lyu, MalPat: Mining
Patterns of Malicious and Benign Android Apps via
Permission-Related APIs",in IEEE Transactions on
Reliability, vol. 67, no. 1, pp. 355-369, March 2018.
[2] K. Xu, Y. Li and R. H. Deng, ICCDetector: ICC-Based
Malware Detection on Android,"inIEEETransactionson
Information Forensics and Security, vol. 11, no 6, pp.
1252-1264, June 2016.
[3] L. Cen, C. S. Gates, L. Si and N. Li, "A Probabilistic
Discriminative Model for Android Malware Detection
with Decompiled Source Code," in IEEE Transactionson
Dependable and Secure Computing, vol. 12, no. 4, pp.
400-412, 1 July-Aug. 2015.
[4] B. Rashidi, C. Fung and E. Bertino,"Android malicious
application detection using support vector machineand
active learning," 13th International Conference on
Network andServiceManagement(CNSM),Tokyo,2017,
pp. 1-9.
[5] N. Peiravian and X. Zhu,"Machine Learning for Android
Malware Detection UsingPermissionandAPICalls"IEEE
25th International Conference on Tools with Arti_cial
[6] https://www.python.org/downloads/ [Accessed on
16/4/2019].
[7] https://www.geeksforgeeks.org/ [Accessed on
21/3/2019].